The Logical Topology and Compactness

Earlier drafts called this the Scott topology and the resulting space a spectral space — sober, \(T_0\), quasi-compact. That is the right description for a subbasis of positive observations and it is the wrong description for this one. \(\HML(\ctx)\) has negation. Definition 16.2 carries it, the namespace logic of \(\chSci\) uses it, and Conjecture 42.1 quantifies over \(\neg\phi\) explicitly. So \(\llb \neg\phi \rrb\) is the complement of \(\llb \phi \rrb\), every subbasic open is also closed, and the space is zero-dimensional. By the Hennessy–Milner property distinct bisimulation classes are separated by some formula, hence by a clopen set, so the space is Hausdorff and its points are closed.

The setting is therefore Stone duality and not domain theory. The specialisation order, which carries the content in a Scott topology, is discrete here and carries none; “sober and \(T_0\)” understates the separation badly. This is not a question of which citation to attach. The two settings answer differently when asked what compactness of a population means. In the Stone setting \(X_\pop\) compact says that \(\terms(\GSLT)/{\bisim}\) restricted to the population is the Stone space of the Boolean algebra of \(\HML(\ctx)\)-definable sets — a profinite limit of finite quotients, each one the population as seen at a bounded modal depth.

That is the same structure §21.5 of \(\chSci\) puts on the hypothesis space, where the ultrametric comes from agreement up to depth and the balls are the levels of the same inverse system. The two parts have been using one object under two descriptions. The logical metric \(d_{\HML}\) of Part Part I is the ultrametric that induces this topology, and the metric balls are exactly the basic clopens.

The latent subspace is “latent” in the sense that it is not constructed deliberately by the agents. It is the topological shadow cast by the population’s collective behavior — the smallest closed set in the logical topology that contains all the agents. An agent inside \(\pop\) cannot directly observe \(X_\pop\); it can only probe it indirectly through experiments.

The proof sketch defined a family of formulae of increasing modal depth, \[\phi_1 = \langle 0 \rangle\top, \quad \phi_2 = \langle 1 \rangle\langle 0 \rangle\top, \quad \phi_3 = \langle 0 \rangle\langle 1 \rangle\langle 0 \rangle\top, \quad \ldots\] with \(0\) and \(1\) labelling the two communities’ states, so that \(\phi_n\) holds of an agent that has completed at least \(n\) oscillation steps. It then claimed that \(\{ \llb \phi_n \rrb \}_{n \in \NN}\) is an open cover of \(X_\pop\) with no finite subcover.

It is not. Satisfying \(\phi_{n+1}\) entails satisfying \(\phi_n\), so the family is decreasing: \(\llb \phi_1 \rrb \supseteq \llb \phi_2 \rrb \supseteq \cdots\). Hence \(\bigcup_{n \le N} \llb \phi_n \rrb = \llb \phi_1 \rrb\) for every \(N\), and an agent that has oscillated more than \(N\) times lies inside that union rather than outside it. There is no failure of finite subcovering and therefore no contradiction with compactness.

The natural repair is to reindex, taking \(\psi_n = \neg\phi_n\) — “has completed fewer than \(n\) steps” — so that the family is increasing. That family is an open cover of \(X_\pop\) precisely when every agent oscillates only finitely often, which is the conclusion. The repair assumes what it sets out to prove.

The other natural repair runs compactness in its finite-intersection form. The \(\llb \phi_n \rrb\) are closed as well as open, by Remark 42.1, and the family has the finite intersection property whenever some agent has oscillated \(N\) times for each \(N\). Compactness then yields \(\bigcap_n \llb \phi_n \rrb \neq \emptyset\) — an agent satisfying every \(\phi_n\), which is to say an infinite oscillator. This is a correct argument and it establishes the opposite of the conjecture. Compactness is a condition guaranteeing that limits are attained, and a perpetual disagreement is exactly such a limit.

The third argument having gone the wrong way is the informative one. Compactness in the logical topology is a richness condition: by Stone duality it says that every finitely satisfiable set of formulae is realised by some point of \(X_\pop\), so the population’s closure omits no consistent type. Nothing about richness makes a process stop.

Termination is not a topological property of the population; it is a budgetary property of the process that revises theories. The framework has the missing hypothesis and it is not in this chapter. Every revision is an assay and every assay is paid for (§21.6 of \(\chSci\)), so if each alternation between \(\phi\) and \(\neg\phi\) costs a bounded-below quantity from a bounded-above endowment, the alternation terminates, and it does so for reasons that have nothing to do with the topology. Compactness would then contribute the second half of the statement rather than the first: it bounds the depth at which the surviving disagreement can be located, and hence the convergence time, once termination is available from the budget.

Two routes out, then. Add the cost hypothesis and prove the conjecture as a statement about priced revision, in which case it belongs in Part Part II and not here. Or keep it topological and weaken the conclusion from “the alternation terminates” to “the alternation is confined to a subspace of bounded logical depth”, which may be provable as stated and is weaker than what the rest of this part uses. The first looks more likely to be true and more useful. Until one of them is done this is a conjecture, and the two results downstream of it — coherence clusters as maximal compact subpopulations, and the reading of non-compactness as permanent disagreement — should be read as depending on it.

Conjecture 42.1 is an existence claim: it says consensus must be reached, without specifying how. A proof of compactness for a specific population subspace is constructive, however: it identifies the finite subcover, which encodes the maximum depth of oscillation and therefore the convergence time. The proof of compactness would be the consensus protocol. This is the sense in which the topology was intended to do computational work, and it is worth stating because it is the payoff that makes the conjecture worth settling.

On the intended reading, a non-compact population may contain irresolvable oscillations — communities that never agree — and non-compactness is the topological signature of fundamental disagreement: not a disagreement that more communication or better arguments could resolve, but one that is structurally permanent. In the context of the massive population, non-compactness at a given scale would mean the cluster at that scale cannot function as a coherent agent. Remark 42.4 is the reason this is stated in the conditional. Under Stone duality non-compactness says the population omits a consistent type, which is a statement about gaps in the population rather than about deadlock within it, and the two readings have not been reconciled.